When we started building Verity, a tool to score the credibility of news text, the default instinct was to reach for a large language model. It would have worked. It also would have been the wrong call.
The task was a well-defined binary classification problem: given a piece of text, is it more consistent with reliable reporting or with misinformation patterns. That kind of problem doesn’t need billions of parameters. It needs good features and a clean training set.
We ended up with TF-IDF vectorization feeding a logistic regression classifier. It trains in seconds, runs in milliseconds, costs nothing per request, and — unlike a black-box model — we can inspect exactly which terms are driving a prediction.
The lesson we keep relearning: the model should match the problem’s actual shape, not the most impressive tool available. An LLM is the right choice for open-ended reasoning. It’s often the wrong choice for a task that a simpler model already solves well.